Deformable Registration with the SIFT points

SIFT points are the features point whose proprities are invariant in scaling, in the change of light, or rotation.

The principal idea of this method is use the correspondance between the points to determiner the transformation, the displacement of point. Then those points are considered as the control points that influence the form of the moving image in the deformation.

The correspondance of the SIFT point in two 2D image,

The number of point-matching pair is limited to increase the correct of the correspondance. I erase the pair whose distance is greater than the average distance of matching.

Then, we use the matching points as the control point for the WarpMethod and LandMarkBase (implemented in Itk).

After getting the result of SIFT for the deformation. The result is not good. I could not able to use SIFT to get the non-rigid registration. I think that SIFT point matching is not enough to the physical law deformation of the region.

Some additional reasons for not using Sift is :
- The obtained SIFT points are not a set of points who determine the shape of the volumn, the form of the region.
- There are some missing in the matching. These missing influence much to the deformation result.
- The 2D SIFT point is 2D. So, it's so complicated to apply to the 3D deformation.

These images are the result of using SIFT with WarpMethod (the coefficience of matching is respective 0.9 and 0.5) and with the landMarkBase (last image) in Itk.

The first image is so bad. The missing match influences much the final result. The second and third images are  only the rigid transfomation. The non-rigid is not obtained.